Suffixscape › Agency Diffusion

Diagnostic 2: Agency Diffusion

The mechanism — two variants

Variant A: the classic passive

"It was determined that the dataset would be retired." Determined by whom? The passive voice makes the agent grammatically optional, and institutional prose exercises the option with remarkable consistency in exactly the sentences where accountability is at stake. The decision arrives pre-made, authorless, as if discovered rather than taken. Note the asymmetry that makes this diagnostic rather than stylistic: the same documents use crisp active voice for achievements ("we shipped", "we achieved") and the agentless passive for harms and reversals ("mistakes were made", "access was restricted"). The grammar tracks the liability.

Variant B: the inanimate cognitive subject

"The model understands user intent." "The system decides which content to show." "The algorithm believes you'll enjoy this." Here agency is not deleted but relocated — into an artefact that cannot hold it. Every such sentence conceals a chain of human choices: who selected the training data, who defined the objective, who set the threshold, who shipped it. Attributing the cognition to the artefact dissolves that chain. This variant matters doubly for AI discourse because it is also the grammar of oracle-positioning: the manifesto's point that AI systems should be treated as mirrors of human editorial choices is, at sentence level, a campaign against this exact construction.

Detection in practice

• Passives: search for was/were/has been/have been/is being + past participle; flag instances with no by-phrase and no agent in the surrounding sentences. Weight decision-verbs (determined, decided, approved, removed, restricted) heavily.
• Cognitive subjects: search for model/system/algorithm/AI/platform/technology/data as subject of understands, knows, believes, thinks, perceives, decides, wants, chooses, learns.
• Apply the attribution test: can the sentence be rewritten with a named human or institutional agent using only information the author plainly had? If yes, the diffusion was a choice.
• Map the active/passive asymmetry across a whole document: which kinds of events get agents, and which don't? The pattern of the pattern is the strongest finding.

Innocent uses (when the hit discharges)

The passive is legitimate when the agent is genuinely unknown, irrelevant, or deliberately de-emphasised for honest reasons (methods sections: "samples were heated"). Cognitive verbs for systems can be defensible shorthand in technical contexts where the community understands the idiom ("the parser expects a newline"). The hit confirms when the construction appears in accountability-bearing sentences — decisions, harms, denials, predictions about people's lives — where a knowable agent has gone missing.

Worked example

“Concerns were raised, and after careful consideration it was concluded that the feature could not be supported. The system will now determine eligibility automatically.”

Four diffusions in two sentences: raised by whom, considered by whom, concluded by whom — and then the remaining agency handed to "the system", which will "determine" eligibility that humans defined, coded, and chose to automate. A reader finishes the passage knowing a decision happened and unable to name a single decider. That unaccountability is manufactured, sentence by sentence, and it is reversible the same way.

Relation to other instruments

Agency diffusion is the central diagnostic — the other three describe its lexical (nominalisation), evidential (inflation), and temporal (flatness) faces. At content level it pairs with Future Modelling (futures without deciders) and Marginalised Voices (the diffused agent and the silenced stakeholder are often the same person, seen from opposite sides).